Skip to content
Open access

OFATr: Optical Flow Attention Transformer for Self-Supervised Monocular Depth Estimation

2026 · IEEE Access · Vol 14, pp. 167071-167090 · 0 citations · 35 references

Abstract

Self-supervised monocular depth estimation has advanced through photometric consistency, yet lightweight networks still infer depth from appearance alone at test time. Such appearance-only inference remains fragile in textureless regions and repetitive structures, where local cues fail to uniquely determine scene geometry. Optical flow between two consecutive frames offers a motion cue, but the cue is spatially unreliable when low texture, small camera motion, or independently moving objects corrupt the flow signal. This paper shows that the cue still improves depth estimation when appearance-conditioned filtering suppresses it in weakly textured regions before it guides self-attention. We propose OFATr (Optical Flow Attention Transformer), a dual-encoder network whose flow-guided self-attention refines appearance features through an auxiliary flow query. At test time, the model uses a reference RGB image and a single-channel optical-flow magnitude map. OFATr combines appearance-conditioned flow gating, dual-query attention, and sinusoidal positional encoding to admit the cue only where appearance indicates reliability. On KITTI, OFATr reaches an absolute relative error of 0.098 and a 1.25-threshold accuracy of 0.899, using 5.2 million parameters including flow preprocessing. It outperforms Lite-Mono-8M and DNA-Depth-B1 with fewer parameters than either, and the margin holds across five seeds. Ablations show that removing the gate returns accuracy to the appearance-only baseline and that standard cross-attention recovers less than half the gain. The model also transfers to Cityscapes without fine-tuning, again with a lower absolute relative error than every baseline. These results indicate that two-frame motion cues resolve appearance ambiguity when they are used to selectively guide the depth representation.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.